Re: [PATCH v4 0/4] Add lightweight memory barriers for coherent memory access

[Alexander Duyck]

On 11/18/2014 03:06 PM, Linus Torvalds wrote:
> On Tue, Nov 18, 2014 at 2:47 PM, Alexander Duyck
> <alexander.h.duyck@xxxxxxxxxx> wrote:
> > The problem is DMA is a broad brush.  There are multiple cases I can think
> > of where DMA does not represent coherent memory.
>
> .. and I already addressed that, in the thing you even included:
>
> > > about what is actually the important issue. All sane memory is
> > > coherent, after all (and if it isn't, you have other issues than
> > > memory ordering).
>
> The thing is, if the DMA isn't coherent, nobody is going to care about
> the memory barriers anyway. You have bigger issues.
>
> And your argument is that "dma" is bigger than this issue. *MY*
> argument is that "coherent" is bigger than this issue. There's tons of
> coherent memory that is not about DMA, the same way that there is DMA
> memory that isn't coherent.
>
> See? The two are 100% equivalent. Except "dma" is just three letters,
> and matches "smp" both visually and in use (SMP memory is "coherent"
> too - yes, you can - and crap architectures do - have incoherent
> caches due to virtual aliases etc, but exactly as with DMA, if you
> have incoherent SMP, you have bigger issues than the barriers).

Actually if anything maybe the crap architectures are a good reason for 
changing the name.  If they can't even do coherent SMP memory then the 
coherent_*mb() could be misleading since they would just be full 
barriers anyway.

> And yes, you could call it "coherent_dma_read_memory_barrier()", and
> it would be very descriptive. It would also drive everybody crazy.

No, I think "dma_wmb__before_coherent_write" would have been much more 
descriptive.  You have to squeeze in that extra underscore somewhere. ;-)

> So I argue for "dma_mb()" pairing with "smp_mb()" from a naming
> standpoint. It just *describes* the problem better. Look at the
> drivers, it's very much about the devices doing DMA to memory, and our
> ordering.
>
> To be even more clear: nobody sane cares about the "coherent" part,
> because only insane horrible crap architectures have incoherent memory
> in the first place, and sane people run away screaming from that
> steaming pile of sh*t.

I think that is part of my reluctance.  I didn't even want it implied 
that the barriers could be used with that kind of stuff.

> Just look at some of the drivers you actually *use* this in. They are
> for intel hardware, they presumably would never even work in the first
> place without cache-coherent DMA. Why do you think that "coherent" is
> so important?
>
>                         Linus

v5 should be up shortly after a quick pass with sed to do the 
find/replace, clean up any whitespace issues, and a quick run through 
some cross compiling scripts just to make sure I didn't screw anything up.

- Alex
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/